Object Category Detection using Audio-visual Cues

Luo Jie1,2, Barbara Caputo1,2, Alon Zweig3,Jorg-Hendrik Bach4, and Jorn Anemuller4

1 IDIAP Research Institute, Centre du Parc, 1920 Martigny, Switzerland2 Swiss Federal Institute of Technology in Lausanne(EPFL), 1015 Lausanne, Switzerland

3 Hebrew university of Jerusalem, 91904 Jerusalem, Israel4 Carl von Ossietzky University Oldenburg, 26111 Oldenburg, Germany

{jluo,bcaputo}@idiap.ch zweiga@cs.huji.ac.il{joerg-hendrik.bach, joern.anemueller}@uni-oldenburg.de

Abstract. Categorization is one of the fundamental building blocks of cognitivesystems. Object categorization has traditionally been addressed in the vision do-main, even though cognitive agents are intrinsically multimodal. Indeed, biologi-cal systems combine several modalities in order to achieve robust categorization.In this paper we propose a multimodal approach to object category detection,using audio and visual information. The auditory channel is modeled on biologi-cally motivated spectral features via a discriminative classifier. The visual channelis modeled by a state of the art part based model. Multimodality is achieved usingtwo fusion schemes, one high level and the other low level. Experiments on sixdifferent object categories, under increasingly difficult conditions, show strengthsand weaknesses of the two approaches, and clearly underline the open challengesfor multimodal category detection.

Key words: Object Categorization, Multimodal Recognition, Audio-visual Fu-sion

1 Introduction

The capability to categorize is a fundamental component of cognitive systems. It canbe considered as the building block of the capability to think itself [1]. Its importancefor artificial systems is widely recognized, as witnessed by a vast literature (see [2, 3]and references therein). Traditionally, categorization has been studied from an unimodalperspective (with some notable exceptions, see [4] and references therein). For instance,during the last five years the computer vision community has attacked the object cat-egorization problem by (a) developing algorithms for detection of specific categorieslike cars, cows, pedestrian and many others [2, 3]; (b) collecting several benchmarkdatabases and promoting benchmark evaluations for assessing progresses in the field.The emerging paradigm from these activities is the so-called ‘part-based approach’,where visual categories are modeled on the basis of local information. This informationis then used to build a learning based algorithm for classification. Both probabilistic anddiscriminative approaches have been used so far with promising results.

Still, an algorithm aiming to work on an autonomous system cannot ignore the in-trinsic multimodal nature of categories, and the multi sensory capabilities of the system.

2 Luo Jie, Barbara Caputo, Alon Zweig, Jorg-Hendrik Bach, and Jorn Anemuller

For instance, we do recognize people on the basis of their visual appearance and theirvoice. Linen can be easily recognized because of its distinctive textural visual and tac-tile properties; and so forth. Biological systems combine information from all the fivesenses, so to achieve robust perception (see [5] and references therein).

In this paper we propose an audio-visual object category detection algorithm. Weconsider categories like vehicles (cars, airplanes), instruments (pianos, guitars) and an-imals (dogs, cows). We assume that the category has been localized, and we focus onhow to integrate together effectively the two modalities. We represent visual informa-tion using a state of the art part based model (section 2, [3]). Audio information isrepresented by a discriminative classifier, trained on biologically motivated spectralfeatures (section 3.1, [6]). Following results from psychophysics, we propose to com-bine the two modalities with a high level fusion scheme that extends previous work onintegration of multiple visual cues (section 3.2, [7]). Our approach is compared withsingle modality classifiers, and with a low level integration approach. Experiments onsix different object categories, with increasing level of difficulty, show the value of ourapproach and clearly underline the existing challenges in this domain (section 3.3).

2 Vision Based Category Detection

In this section we present the chosen vision based category detection algorithm (section2.1) and experiments showing its strengths and weaknesses (section 2.2).

2.1 Visual Category Detection

To learn object models, we use the method described in [3]. The method starts by ex-tracting interest regions using the Kadir & Brady (KB) [8] feature detector. After theirinitial detection, selected regions are cropped from the image and scaled down to 11×11pixel patches, represented using the first 15 DCT (Discrete Cosine Transform) coeffi-cients (not including the DC). To complete the representation, 3 additional dimensionsare concatenated to each feature, corresponding to the x and y image coordinates ofthe patch, and its scale respectively. Therefore each image I is represented using anunordered set F(I) of 18 dimensional vectors. The algorithm learns a generative rela-tional part-based object model, modeling appearance, location and scale. Each part ina specific image Ii corresponds to a patch feature from F(Ii). It is assumed that the ap-pearance of different parts is independent, but this is not the case with the parts’ scaleand location. However, once the object instances are aligned with respect to locationand scale, the assumption of part location and scale independence becomes reasonable.Thus a 3-dimensional hidden variable C = (Cl,Cs), which fixes the location of the ob-ject and its scale, is used. The model’s parameters are discriminatively optimized usingan extended boosting process. For the full derivation of the model and further details,we refer the reader to [3].

2.2 Experiments

We used an extensive dataset of six categories (airplanes, cars, cows, dogs, guitars andpianos). They present different type of challenges: airplanes and cars contain relatively

Object Category Detection using Audio-visual Cues 3

Objects Background

Plan

e Google

Car

Dog

s Road

Cow

SiteG

uita

r Objects

Pian

o

Fig. 1. Sample images from the datasets. Object images appear on the left, background imageson the right.

small variations in scale and location, while cows and dogs have a more flexible ap-pearance and variations in scale and locations. The category images and the back-ground classes were collected from standard benchmark datasets (Caltech Datasets5

and PASCAL Visual Challenge6). For each category we have several correspondingtesting backgrounds, containing natural scenes or various distracting objects. Imagesfrom the six categories and the background groups are shown in Figure 1. Each cat-egory was trained and tested against different backgrounds. Each experiment was re-peated several times, with randomly generated training and test sets. Table 1 presentsthe average results, for different categories and varying backgrounds. These numberscan be compared with those reported in [3], and show that the method delivers state ofthe art performance. We then run some experiments to challenge the algorithm. Namely,we reduced the training set to roughly 1/3 for some categories (airplanes, cars; resultsreported in Figure 2) and, for all categories, we collected new test images containingstrong occlusions, unusual poses and high categorical variability. Exemplar challengingimages are shown in Figure 2. We used the learnt models to classify these challengingimages. These results are also reported in Figure 2. We see that, under these conditions,performance drops significantly for all categories. Indeed, these results seem to indi-cate that the part-based approach might suffer when different categories share similarvisual part (dogs and cows sharing legs, cars and airplanes sharing wheels), or whenthe variability within a single category is very high, as it is for instance for grand pi-anos and upright pianos, or classic and electric guitars. It is worth stressing that theseconsiderations are likely to apply to any part-based visual recognition method. Thus,our multimodal approach for overcoming these issues is of interest for a wide variety ofalgorithms.

5 Available at http://www.vision.caltech.edu/archive.html6 Available at http://www.pascal-network.org/challenges/VOC/

4 Luo Jie, Barbara Caputo, Alon Zweig, Jorg-Hendrik Bach, and Jorn Anemuller

Object Background FNR FPR ERR Background FNR FPR ERRAirplanes Google 2.05 0.81 1.46 Road 1.90 4.96 3.62

Cars Google 5.70 0.41 2.25 Road 11.39 0.73 3.69Cows Site 19.70 2.22 6.91 Road 9.09 0.18 1.14Dogs Site 17.00 13.89 15.53 Road 1.90 0.55 0.91

Guitars (Electrical) Google 9.36 7.59 8.49 Objects 3.69 1.68 2.75Pianos (Grand) Google 18.89 1.41 4.23 Objects 6.67 6.56 6.60

Table 1. Performance of our visual category detection algorithm on six different objects on var-ious background. False Negative Rate(FNR) = num. o f f alse neg.

num. o f pos. instances , False Positive Rate(FPR) =num. o f f alse pos.

num. o f neg. instances and Error Rate(ERR) = num. o f f alse predictiontotal num. o f instances are reported separately.

Object BackgroundVisual

FNR FPR ERRAirplane? Road 26.29 13.89 19.97

Car? Road 38.02 1.91 13.54Cow◦ Site 78.33 - 78.33Dog◦ Site 27.00 - 27.00

Piano† Google 58.48 - 58.48Guitar† Google 16.00 - 16.00

?: reduced number of training samples;◦: learnt models of cows & dogs to detect new testimages with occlusions and strange pose;†: learnt models of grand pianos & electricalguitars to detect upright piano and classical guitarrespectively.

(a) Hard Dog Examples;

(b) Hard Cow Examples;

(c) Four-legged animals;

(d) Upright Piano

Fig. 2. Performance of the visual category detection algorithm on various difficulty examples.Some exemplary images are shown on the right of the table.

3 Audio-visual Category Detection

This section presents our multi-modal approach to object category detection. We beginby illustrating the sound classification method used (section 3.1). We then illustrate ourintegration method (section 3.2) and show with an extensive experimental evaluationthe effectiveness of our approach (section 3.3).

3.1 Audio Category Detection

Real-world audio data is characterized in particular by two properties, spectral charac-teristics and modulation characteristics. Spectral characteristics are obtained by decom-posing the signal into different spectral bands, typically using “Bark-scaled” frequencybands that approximate the spectral resolution of the human ear. Here, we use 17 Barkbands ranging from about 50 Hz to 3800 Hz. Within each spectral band, information

Object Category Detection using Audio-visual Cues 5

about the signal is encoded in changes of spectral energy across time, so-called ampli-tude modulations (2 Hz to 30 Hz). Grouping both properties in a single diagram, we ob-tain the “amplitude modulation spectrogram” (AMS, [9]), a 3-dimensional signal repre-sentation with dimensions time, (spectral) frequency and modulation frequency. Each 1slong temporal window is represented by 17×29 = 493 points in frequency/modulation-frequency space. Audio category detection [6] is performed by linear SVM classifica-tion based on a subset of the 493 AMS input features, trained to discriminate betweenaudio samples containing only background noise (e.g., street) and samples containingan audio category object (e.g., dog) embedded in background noise at different signal-to-noise ratios (from +20 dB to -20 dB).

3.2 Audio-visual Category Detection

This section provides a short description of our cue integration scheme. Many cue in-tegration methods have been presented in the literature so far. For instance, one candivide them in low level and high level integration, where the emphasis is on the level atwhich integration happens [4]. In low level integration, information is combined beforeany use of classifiers or experts. In high level approaches, integration is accomplishedby an ensemble of experts or classifiers; on each prediction, a classifier provides a harddecision, an expert provides an opinion. In this paper, we will investigate methods fromboth approaches.

High Level Integration There are several methods for fusing multiple classifiers atthe decision level [10], such as voting, sum-, product-rule, etc. However, voting couldnot be easily applied on our setup, since it requires an odd number of classifiers for atwo class problem, and more for a multi-class problem. Here we use an extension of thediscriminative accumulation scheme (DAS) [7]. The basic idea is to consider the marginoutputs of any discriminative classifiers (e.g. AdaBoost and SVMs) as a measure of theconfidence of the decision for each class, and accumulate all the outputs obtained forvarious cues with a linear function. The binary class version of the algorithm could bedescribed into two steps:

1. Margin-based classifiers: These are a class of learning algorithms which take asinput binary labeled training examples (x1, y1), . . . , (xm, ym) with xi ∈ χ and yi ∈

{−1,+1}. Data are used to generate a real-valued function or hypothesis f : χ→<,with f belonging to some hypothesis space F. The margin of an example x withrespect to f is f (x), which is determined by minimizing: 1

m∑m

i=1 L(yi f (x)), for someloss function: L : <→ [0,∞] Different choices of the loss function L and differentalgorithms for minimizing the equation over some hypothesis space lead to variouswell studied learning algorithms such as Adaboost and SVMs.

2. Discriminative Accumulation: After all the margins are collected { f pj }

Pp=1, for all

the P cues, the data x is classified using their linear combination:

J = sgn

P∑p=1

wp f pj (xp)

.

6 Luo Jie, Barbara Caputo, Alon Zweig, Jorg-Hendrik Bach, and Jorn Anemuller

The original DAS method considered only SVM as experts, used multiple visualcues for training and determined the weighting coefficients via cross validation.Here we generalize the approach in many respects: we take two different largemargin classifiers (SVM and AdaBoost) as experts, we train each expert on a dif-ferent modality, and we determine the weights {wp}

Pp=1 by training a single-layer

artificial neural network (ANN) on a validation set.

A drawback of the original DAS algorithm is that the accumulation function is lin-ear, thus the method is not able to adapt to the special characteristics of the model. Forexample, one sensor might be suddenly affected by noise, or detect a novel input. Herewe will assume that if one sensor is very confident about the presence of a category (i.e.margin above a certain threshold), it is highly probable that this sensor is correct. Wethus introduce a threshold before the accumulation, so that if the margin output valueof one classifier is larger than the threshold, we will take it directly as the decision.

Low Level Integration The low level fusion is also known as feature level fusion.Features extracted from data provided by different sensors are combined. In case of au-dio and visual feature vectors, the simple concatenation technique could be employed,where a new feature vector can be built by concatenating two feature vectors together.There are a few drawbacks to this approach: the dimensionality of the resulting featurevector is increased, and the two separate feature vectors must be available at the sametime (synchronous acquisition). Due to the second problem, the high level integrationis usually preferred in the literature for audio-visual fusion [4].

The visual feature vectors are built by concatenating all the P feature descriptors.Each feature consists of a 20-dimensional vector including [3]: the 18 dimensional vec-tor representing each image (see section 2.1), plus a normalized mean of the featureand a normalized logarithm of the feature variance. The training set is then normalizedto have unit variance in all dimensions, and the standard deviations are stored in or-der to allow for identical scaling of the test data. Finally, the visual feature vector isconcatenated with audio feature vectors, and a linear SVM is trained for the detection.

3.3 Experiments

Experimental Setup We evaluated our multi-modal approaches with three series ofexperiments. Our audio dataset contains a large number of audio clips, manually col-lected from the internet, corresponding to the six visual categories as well as someother objects. The audio background noise class contains recordings of road traffic andpedestrian zone noise. All audio models were trained with the same background noisebut different object sounds, on several combinations of training and test sets, randomlygenerated. Then each audio file (object/background) was randomly associated with animage (object/background) without repetitions. Audio and visual data were collectedseparately, and their association was somehow arbitrary. Thus, we repeated the experi-ments at least 1,000 times, for each setup, so to prevent “lucky” cases. We compared ourresult on fusion with those obtained by single cues, reporting average results. For DAS,we experimented with the linear and non-linear (i.e. with an additional threshold forhigh-confidences input) approaches. However, we did not find significant differencesbetween them. Thus we only report results obtained using the linear method.

Object Category Detection using Audio-visual Cues 7

Object BackgroundAudio Visual Fusion

FNR FPR ERR FNR FPR ERR FNR FPR ERR

Hig

h-le

vel

Airplane Road 7.53 11.41 9.71 1.55 2.99 2.36 1.36 1.52 1.45Cars Road 6.73 3.22 4.20 11.33 0.71 3.67 1.51 0.70 0.93Cows Site 4.62 0.47 1.49 19.97 2.21 6.97 2.40 0.62 1.10Dog Site 2.09 0.36 1.27 16.78 13.79 15.37 0.36 0.62 0.48

Piano Google 7.98 0.09 1.36 18.87 1.41 4.21 1.75 0.31 0.54Guitar Google 7.53 0.09 3.16 9.32 7.57 8.29 0.48 0.30 0.38

Low

-lev

el

Airplane Road 8.12 6.19 7.03 4.63 9.57 7.40 3.49 2.31 2.83Cars Road 7.52 1.35 3.06 8.27 1.46 3.35 3.67 0.19 1.16Cows Site 5.37 0.03 1.47 16.74 6.44 9.20 5.13 0.00 1.39Dog Site 1.96 0.01 0.99 16.93 22.91 19.76 1.83 0.01 0.97

Piano Google 8.50 0.00 1.37 21.56 0.81 4.16 7.50 0.00 1.21Guitar Google 4.15 0.06 2.15 18.5 1.45 10.17 3.80 0.02 1.96

Table 2. Results of each separate audio and visual cues and detection performance of both thehigh- and low-level integration scheme on six different objects.

Experiments with Clean Data Table 2 reports the FNR., FPR., and ERR. for differentobjects, using high- and low-level fusion schemes. For each object, we performed ex-periments using various backgrounds. For space reasons we report here only a represen-tative subset. Results show clearly that, for all objects and both fusion schemes, recog-nition improves significantly when using multiple cues, as opposed to single modalities.Regarding the comparison between the two fusion approaches, it is important to stressthat, due to the different classification algorithms, and the differences in statistics inthe training data for the different classes, it is not straightforward how to compare theperformance of the two fusion schemes. Still, the high-level scheme seems to obtainoverall lower error rates, compared to the low-level approach.

Experiments with Difficult/Noisy Data We tested the robustness of our system withrespect to noisy cues or difficult sensory inputs. First, we showed the effects of includingaudio cues for improving the system performance when there are not enough trainingexamples (see section 2.2); results are reported in Table 3. Then, we used the modelstrained on clean data and test them on various difficult images (see section 2.2). Theseresults are also shown in Table 3. Finally, the systems were tested against noisy audioinputs. The performance of the audio classifier was deliberately decreased by addingvarying amount of street noise on the test object audio (SNR ∈ [−20db, 20db]), andincluding a varying amount of audio files generated by other objects 7 as part of thetest background noise (from 25% to 1% of the total number of testing backgroundsamples). Average results on two selected examples, car (less training examples, roadbackground) and dog (site background), are shown in Figure 3.

With respect to the high-level and low-level fusion methods, we see that the lowlevel approach seems to be more robust to noise (Table 3 and Figure 3). This might be

7 Sounds of other animals, e.g. bear, horse, in case of experiments on cows and dogs, and soundsof artificial objects, e.g. phone, helicopters, in case of vehicles and instruments.

8 Luo Jie, Barbara Caputo, Alon Zweig, Jorg-Hendrik Bach, and Jorn Anemuller

Object BackgroundAudio Visual Fusion

FNR FPR ERR FNR FPR ERR FNR FPR ERR

Hig

h. Airplane (Less) Road 8.72 11.90 10.33 26.24 13.89 19.97 6.17 7.62 6.90Cars (Less) Road 6.80 3.31 4.46 38.02 1.91 13.54 3.69 1.76 2.38

Low. Airplane (Less) Road 5.36 9.65 7.76 3.78 29.16 18.01 2.76 6.96 5.12

Cars (Less) Road 7.50 1.57 3.21 14.08 10.01 11.15 4.83 0.50 1.70

Table 3. Performance of the multimodal system suffering from less visual training examples.

Object BackgroundHigh-level Low-level

Audio Visual Fusion Audio Visual FusionCows (Hard) Site 7.24 78.33 7.27 7.54 77.02 7.92Dog (Hard) Site 2.69 27.00 2.32 2.56 29.24 2.54

Piano (Upright) Google 7.21 58.48 16.35 8.43 39.39 8.41Guitar(Classical) Google 7.14 16.00 6.40 6.04 29.80 5.93

Table 4. Performance of the multimodal system in presence of difficult test examples with oc-clusions, unusual poses and high categorical variability. Since background images were not usedduring test, only the false negative error rates (FNR) are reported.

due to the nature of the two algorithms, as for the low level fusion method the error rateis linked to the lower error rate between the two cues. The high-level fusion scheme in-stead weights the confidence estimates from the two sensory channels with coefficientslearned during training. Thus, if one modality was weighted strongly during training,but is very noisy during test, the high-level scheme will suffer from that.

Experiments with Missing Audio An important issue when working on multimodalinformation processing is the synchronicity of the two modalities, i.e. both audio andvisual input must be perceived together by the system. However, unlike the multimodalperson authentication scenario [4], in real-world cases the two inputs may not alwaysbe synchronized, e.g. a dog might be quiet. We tested our system in the case wheresome of the object samples were not accompanied by audio. To simplify the problem,we only considered the case where roughly 50% of the object samples are “silent”. Weconsidered three different ways to tackle the problem (see Figure 4, caption), and weoptimized our system using a validation set under the same setup. Figure 4 reports theaverage results on the categories car (road background) and dog (site background). Wecan see that the performance still improves significantly when the missing audio inputswere represented using zero values (roughly the same as results reported in Table 2). Forthe other two scenarios, the performance on cars still grows, while the performance ondogs drops because the system was biased toward the audio classifier when the visualclassifier did not have high accuracy. However, the system was always better than usingthe visual algorithm alone.

Object Category Detection using Audio-visual Cues 9

SNR!20 SNR!15 SNR!10 SNR!5 SNR0 SNR+5 SNR+10 SNR+15 SNR+200

5

10

15

20

25

30

Noise on Audio Data

Erro

r Rat

e [%

]

Visual Audio Optimize Clean

SNR!20 SNR!15 SNR!10 SNR!5 SNR0 SNR+5 SNR+10 SNR+15 SNR+200

5

10

15

20

25

30

Noise on Audio Data

Erro

r Rat

e [%

]

Visual Audio Fusion

(a) Cars, High-level Fusion. (b) Cars, Low-level Fusion.

SNR!20 SNR!15 SNR!10 SNR!5 SNR0 SNR+5 SNR+10 SNR+15 SNR+200

5

10

15

20

25

30

35

40

45

50

Noise on Audio Data

Erro

r Rat

e [%

]

Visual Audio Optimize Clean SNR+10

SNR!20 SNR!15 SNR!10 SNR!5 SNR0 SNR+5 SNR+10 SNR+15 SNR+200

5

10

15

20

25

30

35

40

45

50

Noise on Audio Data

Erro

r Rat

e [%

]

Visual Audio Fusion

(c) Dogs, High-level Fusion. (d) Dogs, Low-level Fusion.

Fig. 3. Performance of the multimodal system in the presence of different level of corruptedaudio inputs. For high-level fusion, the accumulating weights were found using different criteria:the weights were determined using clean audio and visual data through previous experiments(Clean), determined using data at current test noisy level (Optimize), or determined using data atfixed noisy level (e.g. SNR+10).

4 Discussion and Conclusions

This paper presented a multimodal approach to object category detection. We consid-ered audio and visual cues, and we proposed two alternative fusion schemes, one high-level and the other low-level. We showed with extensive experiments that using multiplemodalities for categorization leads to higher performance and robustness, compared touni-modal approaches.

This work can be developed in many ways. Our experiments show that the high-level approach might suffer in case of noisy data. This could be addressed by usingadaptive weights, related to the confidence of the prediction for each modality. Also, weestimate confidences using the distance from the separating hyperplane, but other solu-tions should be explored. We also plan to extend our model to a hierarchical represen-tation as in [11]. Finally, these experiments should be repeated on original audio-visual

10 Luo Jie, Barbara Caputo, Alon Zweig, Jorg-Hendrik Bach, and Jorn Anemuller

Car Dog0

2

4

6

8

10

12

14

16

18

20

Object

Erro

r Rat

e [%

]

Audio Visual Zero Random Background

Three ways for representing the missing audio input:Zero: the input confidences of the audio classifier equalzero, if the audio is missing; thus only the visual classi-fier will be considered.Random: the input confidences of the audio classifierequal randomly generated numbers with a zero meanand standard deviation equals one, if the audio is miss-ing;Background: based on the assumption that the environ-mental noise was always presented, the test images wereassociated with a random background audio if the audioinput is missing.

Fig. 4. Performance of the multimodal system under asynchronous test conditions.

data, so to better address the issues of synchronicity and sound-visual localization. Fu-ture work will focus on these issues.

Acknowledgments This work was sponsored by the EU integrated project DIRAC(Detection and Identification of Rare Audio-visual Cues, http://www.diracproject.org/)IST-027787. The support is gratefully acknowledged.

References

1. Pfeifer, R., Bongard, J.: How the body shapes the way we think. MIT Press (2006)2. Fergus, R., Perona, P., Zisserman, A.: Weakly supervised scale-invariant learning of models

for visual recognition. Int. J. Comput. Vision 71(3) (2007) 273–3033. Bar-Hillel, A., Weinshall, D.: Efficient learning of relational object class models. Int. J.

Comput. Vision (2007) in press.4. Sanderson, C., Paliwal, K.K.: Identity verification using speech and face information. Digital

Signal Processing 14(5) (2004) 449–4805. Burr, D., Alais, D.: Combining visual and auditory information. Progress in Brain Research

155 (2006) 243–2586. Schmidt, D., Anemuller, J.: Acoustic feature selection for speech detection based on ampli-

tude modulation spectrograms. In: 33rd German Annual Conference on Acoustics. (2007)7. Nilsback, M.E., Caputo, B.: Cue integration through discriminative accumulation. In: IEEE

Computer Society Conference on Computer Vision and Pattern Recognition. (2004) 578–5858. Kadir, T., Brady, M.: Saliency, scale and image description. Int. J. Comput. Vision 45(2)

(2001) 83–1059. Kollmeier, B., Koch, R.: Speech enhancement based on physiological and psychoacoustical

models of modulation perception and binaural interaction. J. Acoust. Soc. Am. 95(3) (1994)1593–1602

10. Polikar, R.: Ensemble based systems in decision making. IEEE Circuits and Systems Mag.6(3) (2006) 21–45

11. Zweig, A., Weinshall, D.: Exploiting object hierarchy: Combining models from differentcategory levels. In: IEEE 11th International Conference on Computer Vision. (2007) 1–8